README.md

---
task_categories:
  - image-to-text
pretty_name: ATR benchmark
size_categories:
  - n<1K
language:
- fr
- la
- en
- no
- ar
- zh
- de
tags:
- atr
- htr
- ocr
dataset_info:
  features:
  - name: dataset
    dtype: string
  - name: image
    dtype: image
  - name: arkindex_id
    dtype: string
  - name: text
    dtype: string
  - name: language
    dtype: string
  - name: level
    dtype: string
  splits:
  - name: test
    num_bytes: 145153440.0
    num_examples: 83
  download_size: 144199823
  dataset_size: 145153440.0
configs:
- config_name: default
  data_files:
  - split: test
    path: data/test-*
---

# ATR benchmark - Page/paragraph level

## Dataset Description

- **Homepage:** [ATR benchmark](https://huggingface.co/datasets/Teklia/ATR-benchmark)
- **Point of Contact:** [TEKLIA](https://teklia.com)

## Dataset Summary

The ATR benchmark dataset is a multilingual dataset that includes 83 document images, at page or paragraph level. This dataset has been designed to test ATR models and combines data from several public datasets:
- [BnL Historical Newspapers](https://data.bnl.lu/data/historical-newspapers/)
- [CASIA-HWDB2](https://nlpr.ia.ac.cn/databases/handwriting/Offline_database.html)
- [DIY History - Social Justice](http://diyhistory.lib.uiowa.edu/)
- FINLAM - Historical Newspapers
- [Horae - Books of hours](https://github.com/oriflamms/HORAE)
- [NorHand v3](https://zenodo.org/records/10255840)
- [Marius PELLET](https://europeana.transcribathon.eu/documents/story/?story=121795)
- [RASM](http://www.primaresearch.org/RASM2019/)

Images are in their original size.

### Split

| dataset                        | images | language         | 
| ------------------------------ | ------:| ---------------- |
| BnL Historical Newspapers      |      3 | German           |
| CASIA-HWDB2                    |     10 | Chinese          |
| DIY History - Social Justice   |     20 | English          |
| FINLAM - Historical Newspapers |     10 | English / French |
| Horae - Books of hours         |     10 | Latin            |
| NorHand v3                     |     10 | Norwegian        |
| Marius PELLET                  |     10 | French           |
| RASM                           |     10 | Arabic           |

## Dataset Structure

### Data Instances

```
{
  'image': <PIL.JpegImagePlugin.JpegImageFile image mode=RGB size={img size} at 0x1A800E8E190,
  'text': '{transcription}'
}
```

### Data Fields

- `image`: a PIL.Image.Image object containing the image. Note that when accessing the image column (using dataset[0]["image"]), the image file is automatically decoded. Decoding of a large number of image files might take a significant amount of time. Thus it is important to first query the sample index before the "image" column, i.e. dataset[0]["image"] should always be preferred over dataset["image"][0].
- `dataset`: the name of the original dataset.
- `arkindex_id`: Arkindex element id corresponding to the current page or paragraph.
- `text`: the label transcription of the image.
- `language`: language of text on the image.
- `level`: full document page or a single paragraph.